• Contents
• SiamFC Description
• Model Architecture
• Dataset
• Environmental requirements
• quick start
• Script description
• • Script and sample code
  • Script parameters
  • Training process
  • • Training
  • Evaluation process
  • Ascend310 infer
• Model description
• • performance
  • • Evaluate performance
  • performance
  • • Inference on Ascend310 Performance

Contents

• SiamFC Description
• Model Architecture
• Dataset
• Environmental requirements
• Quick Start
• Script Description
  • Script and Sample Code
  • Script Parameters
  • Training Process
    • Training
  • Evaluation Process
    • Evaluation
  • Ascend310 infer
• Model Description
  • Performance
    • Evaluation Performance

SiamFC Description

Siamfc proposes a new full convolution twin network as the basic tracking algorithm, which is trained end-to-end on ilsvrc15 target tracking video data set. Our tracker exceeds the real-time requirement in frame rate. Although it is very simple, it achieves the best performance on multiple benchmarks.

paper Luca Bertinetto Jack Valmadre Jo˜ao F. Henriques Andrea Vedaldi Philip H. S. Torr
Department of Engineering Science, University of Oxford

Model Architecture

Siamfc first uses full convolution alexnet for feature extraction online and offline, and uses twin network to train the template and background respectively. On line, after getting the box of the first frame, it carries out centrrop, and then loads checkpoint to track the subsequent frames. In order to find the box, it needs to carry out a series of penalties on the score graph, Finally, the final prediction point is obtained by twice trilinear interpolation.

Dataset

used Dataset :ILSVRC2015-VID

• Dataset size : 85GB ,total 30 type
  • Training set: a total of 3862 videos and their corresponding frame pictures and box positions
  • Verification set: 555 videos and corresponding pictures and box locations
  • Test set: a total of 973 videos and corresponding pictures and box locations
• Data format: the image is in hwC format, the box position includes the coordinates of the lower left corner and the upper right corner, the format is XML, and the XML needs to be parsed

Environmental requirements

• Hardware :(Ascend)
  • Prepare Ascend processor to build hardware environment
• frame:
  • Mindspore
• For details, please refer to the following resources:
  • MindSpore course
  • MindSpore Python API
• more API
  • got10k toolkit
  • opencv
  • lmdb

quick start

After installing mindspree through the official website, you can follow the following steps to train and evaluate:

• Run the python script to preprocess the data set

  python src/create_dataset_ILSVRC.py -d data_dir -o output_dir

• Run Python script to create LMDB

  python src/create_lmdb.py -d data_dir -o output_dir

  for example：
  data_dir = '/data/VID/ILSVRC_VID_CURATION_train'
  output_dir = '/data/VID/ILSVRC_VID_CURATION_train.lmdb'

  Remarks:The encrypted pathname is used as the index.Therefore,you cannot change the location of the dataset
  after creating it, because you need to find the corresponding image according to the index.

• Run the script for training

  bash run_standalone_train_ascend.sh [Device_ID] [Dataset_path]
  Remarks:For the training set position after preprocessing

• more

  This example is single card training.

• Run the script for evaluation

  python eval.py,need got10k toolkit,the dataset is OTB2013(50) or OTB2015(100)

Script description

Script and sample code

    ├── SiamFC
        ├── README.md                    // Notes on siamfc
        ├── ascend310_infer             // Implementation inference script on ascend310
        │   ├──inc                      //Head file
        │   ├──src                      //Main.cc and utils.cc file
        │   ├──build.sh                 //Build file
        │   ├──CMakeLists.txt           //Required library files
        ├── scripts
        │   ├──ma-pre-start.sh          // Create environment before modelarts training
        │   ├──run_standalone_train_ascend.sh             // Single card training in ascend
        │   ├──run_distribution_ascend.sh          // Multi card distributed training in ascend
        │   ├──run_infer_310.sh         //310infer scripts
        ├── src
        │   ├──alexnet.py             // Create dataset
        │   ├──config.py              // Alexnet architecture
        │   ├──custom_transforms.py   //Data set processing
        │   ├──dataset.py            //GeneratorDataset
        │   ├──Groupconv.py        //Mindpore does not support group convolution at present. This is an alternative
        │   ├──lr_generator.py       //Dynamic learning rate
        │   ├──tracker.py           //Trace script
        │   ├──utils.py             // utils
        │   ├──create_dataset_ILSVRC.py     // Create dataset
        │   ├──create_lmdb.py               //Create LMDB
        ├── train.py               // Training script
        ├── eval.py               //  Evaluation script

Script parameters

python train.py and config.py The main parameters are as follows:

• data_path：An absolutely complete path to training and evaluation data sets.
• epoch_size：Total training rounds
• batch_size：Training batch size.
• image_height：The image height is used as the model input.
• image_width：The image width is used as the model input.
• exemplar_size：Template size
• instance_size：Sample size.
• lr：Learning rate.
• frame_range：Select the frame interval of the template and sample.
• response_scale：Scaling factor of score chart.

Training process

Training

• Running in ascend processor environment

  python train.py  --device_id=${DEVICE_ID} --data_path=${DATASET_PATH}

• After training, the loss value is as follows:

  grep "loss is " log
  epoch: 1 step: 1, loss is 1.14123213
  ...
  epoch: 1 step: 1536, loss is 0.5234123
  epoch: 1 step: 1537, loss is 0.4523326
  epoch: 1 step: 1538, loss is 0.6235748
 ...

• Model checkpoints are saved in the current directory.

• After training, the loss value is as follows:

  grep "loss is " log:
  epoch: 30 step: 1, loss is 0.12534634
  ...
  epoch: 30 step: 1560, loss is 0.2364573
  epoch: 30 step: 1561, loss is 0.156347
  epoch: 30 step: 1561, loss is 0.173423

Evaluation process

Check the checkpoint path used for evaluation before running the following command.

• Running in ascend processor environment

  python eval.py  --device_id=${DEVICE_ID} --model_path=${MODEL_PATH}

The results were as follows:

  SiamFC_159_50_6650.ckpt -prec_score:0.777 -succ_score:0.589 _succ_rate:0.754

Ascend310 infer

Check the checkpoint path used for evaluation before running the following command.

Run this reference scripts need two different MINDIR

  python export.py  --device_id=${DEVICE_ID} --model_path=${MODEL_PATH} --file_name_export1=${SAVE_MODEL_PATH1} --file_name_export2=${SAVE_MODEL_PATH2} --file_name=${FILE_FORMAT} --device_target=${DEVICE_TARGET}

• Running in ascend310 device processor environment

   bash run_infer_310.sh [MODEL_PATH1] [MODEL_PATH2] [DATASET_PATH] [CODE_PATH] [DEVICE_TARGET] [DEVICE_ID]

Model description

performance

Evaluate performance

parameter	Ascend
resources	Ascend 910；CPU 2.60GHz, 192core；memory：755G
Upload date	2021.5.20
mindspore version	mindspore1.2.0
training parameter	epoch=50,step=6650,batch_size=8,lr_init=1e-2,lr_endl=1e-5
optimizer	SGD optimizer，momentum=0.0,weight_decay=0.0
loss function	BCEWithLogits
training speed	epoch time：285693.557 ms per step time :42.961 ms
total time	about 5 hours
Script URL	https://gitee.com/mindspore/models/tree/master/research/cv/SiamFC
Random number seed	set_seed = 1234

performance

Inference on Ascend310 Performance

parameter	Ascend
Model Version	SiamFC
Upload date	2021.11.1
mindspore version	mindspore1.3.0
Dataset	OTB2013
total time	about 5 minutes
outputs	probability
Accuracy	prec_score:0.779 -succ_score:0.588 _succ_rate:0.756